Modeling Strong and Human-Like Gameplay with KL-Regularized Search